What Auwai is
Auwai is a deterministic coverage layer for a codebase. It indexes first-party files into stable chunks, builds a partition tree, plans full-read segments, and records what has been read. The model reads bounded segments and writes findings to disk. The database keeps the map and the coverage state.
Auwai is not a model, an embedding store, or a source editor. The core runs on the Python standard library. Indexing makes no network call and no model call. The optional harness integrations add the agent surface, not the index logic.
System at a glance
The core is deterministic and offline. The optional knowledge service and the harness plugins sit outside the index and planning path.
What it does
- Indexes first-party source, tests, documentation, configuration, and infrastructure.
- Excludes vendor, generated, build, cache, and lock files with a reason.
- Builds a partition tree from the path and module layout.
- Plans a full read from the usable model context, not a fixed percentage of it.
- Stores findings, reads, coverage, plans, and segments in SQLite.
- Serves a deterministic label and FTS index for retrieval.
- Watches the tree, refreshes changed files, and audits drift.
- Blocks raw content reads of indexed first-party files in the Hermes hook.
- Treats a workspace as the root of the same partition tree.
What it does not do
- It does not decide whether code is correct.
- It does not replace tests, review, or source control.
- It does not read every file into one context.
- It does not require embeddings or a reranker.
- It does not edit source files.
Current status
The coverage architecture is the current main implementation.
The full-read path has unit and integration coverage.
The quality run is task-scoped; a full-window segment is not measured yet.